The real cost of an AI agent

We've watched a lot of agencies sell "AI agents" for $5K–$25K, then go quiet six months later. The build looked fine. The demo worked. The model bill didn't.

There are three things that cost money on every AI workflow:

1. Build time — already covered by the build fee.

2. LLM API calls — variable, recurring. GPT-4o is $2.50 per million input tokens and $10 per million output. Claude Sonnet 4.6 is $3 in / $15 out. A customer support bot doing 800 conversations a month at 2,500 tokens in / 400 out per turn racks up real money fast — easily $30–$150/mo on Sonnet, more on Opus.

3. Third-party API calls — VIN lookups, telephony, transcription, web scraping, vector storage. Often more than the LLM. A voice agent stack (Twilio + transcription + TTS + LLM) can cross $0.50/minute on its own.

The pattern we see in failing AI engagements is the agency builds the workflow, hands it over, and tries to fold the recurring cost into a "$200/mo support retainer." That covers the agency's time. It doesn't cover $400/mo of API spend nobody warned the client about.

What we do instead: - Three explicit lines on every AI quote: build (one-time), Care (recurring, our time), API (recurring, pass-through). - Care tiers scale with workflow complexity ($199 / $399 / $899/mo). A voice agent doesn't get the FAQ-bot Care tier. - API costs are always pass-through. Client uses their own keys, or we manage and bill at-cost + 15% admin. - Volume sensitivity is part of the discovery call. We model what happens at 100/500/5,000 requests/mo before signing anything.

If you're shopping for an AI workflow and the agency hasn't told you what it costs to run, that's not because it's small. It's because they don't know yet, and they're hoping you won't either.

You can run the math yourself in 90 seconds at our free cost estimator. Bring the result to whoever you're talking to and watch the conversation get useful.